Associative Memory N-Gram Classification for Large Hypervectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The computational difficulty of implementing hyperdimensional computing (HDC) N-gram operations, particularly for large hypervectors, on standard CPUs, and the inefficiency of existing in-memory solutions like those described by Karunaratne et al., which require approximations and exponential complexity.
Innovation Solution
Utilizing an associative processing unit (APU) to perform XNOR, permute, and add operations on hyperdimensional vectors stored in an associative memory array, enabling efficient calculation of N-grams without approximating the HDC equation, allowing for linear operations in N.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If standard CPUs are used to implement hyperdimensional computing N-gram operations, then the system can process data, but the computational difficulty and time consumption increase significantly for large hypervectors
Solution Approach 1:
The patent replaces standard CPU-based mechanical computation with an optical computing system that uses light propagation and optical components (lenses, mirrors, detectors) to perform hyperdimensional computing operations. This substitution enables parallel processing of large hypervectors through optical interference patterns, achieving significantly faster N-gram classification without the time consumption inherent in sequential CPU operations.
Solution Approach 2:
The patent transitions from sequential processing in the temporal dimension to parallel processing in the spatial dimension by encoding hypervectors as optical patterns. Multiple N-gram operations are performed simultaneously across different spatial locations using optical interference, effectively adding a spatial dimension to computation and enabling parallel processing of large datasets.
2Productivity
If in-memory computing solutions like Karunaratne et al. are used, then processing can be performed, but the solutions require approximations and have exponential complexity
Solution Approach 1:
The patent extracts the essential computational function from complex in-memory systems by using optical interference patterns to directly compute N-gram representations. This extraction eliminates the need for approximations and exponential complexity by using physical optical principles to perform exact computations through light wave interactions.
Solution Approach 2:
The patent uses optical copying of hypervector patterns through light propagation. Instead of manipulating data in memory, the system creates optical copies of hypervectors as interference patterns that can be processed simultaneously. This copying approach enables parallel computation without the complexity inherent in in-memory manipulation methods.
3Measurement precision
If large hypervectors are processed, then accurate language classification is achieved, but the computational operations become intractable on conventional systems
Solution Approach 1:
The patent replaces conventional mechanical computation with optical computing to handle large hypervectors. Optical interference patterns enable the system to process high-dimensional data maintain classification accuracy while achieving computational tractability through parallel optical operations that conventional systems cannot perform.
Solution Approach 2:
The patent segments the hypervector processing into parallel optical operations. By dividing the computational task into simultaneous optical interference patterns across multiple spatial locations, the system maintains accuracy for large hypervectors while making computation tractable through parallel processing.
Data Source
AI summary
A system for N-gram classification in a field of interest via hyperdimensional computing includes an associative memory array and a controller. The associative memory array stores hyperdimensional vectors in rows of the array. The hyperdimensional vectors represent symbols in the field of interest and the array includes bit-line processors along portions of bit-lines of the array. The controller activates rows of the array to perform XNOR, permute, and add operations on the hyperdimensional vectors with the bit-line processors, to encode N-grams, having N symbols therein, to generate fingerprints of a portion of the field of interest from the N-grams, to store the fingerprints within the associative memory array, and to match an input sequence to one of the stored fingerprints.